Sensible-heat polynyas, i.e. persistent polynyas produced thermodynamically, can form across multiple years where relatively warm water flows out from the sub-ice-shelf cavity toward ice shelf fronts, typically directed by channels carved into the ice-shelf base. Polynyas represent an important component of the interactions between ice shelves, channel water outflows, and ambient ocean properties and circulation. However, time-resolved investigations of persistent polynyas are sparse across Antarctica, due to a dearth of field observations and the complexity of identifying their extent and evolution in remotely sensed images. The challenge, particularly along the coastal margin of the West Antarctic Ice Sheet is not one of a paucity of data, but one of a plethora of pixels lacking classification; typically, an expert is required to label pixels belonging to ice shelf, iceberg, sea ice, open ocean, polynya, or a mixture. Here, we engineer a physics-featurized convolutional neural network with an approach that allows us to predict pixel classifications even in the case where expertly labeled data are sparse. We demonstrate the potential of physics-featurized neural networks to robustly automate complex classification tasks, such as pixel-based labeling of remote sensing imagery, that are often a necessary stepping stone to answering more specific scientific questions across the cryosphere.